Multimodal Biometric Security using Evolutionary Computation · 2016-09-09 · Multimodal Biometric Security using Evolutionary Computation 2 1.1 Introduction to Biometric Systems

Multimodal Biometric Security using Evolutionary Computation

A thesis submitted to

Department of Computer Science & Software Engineering

for the Partial Fulfillment of the Requirement of

MS (CS) Degree

By

Qurrat ul Ain

639-FBAS/MSCS/F10

Supervised by

Dr. Ayyaz Hussain

Department of Computer Science and Software Engineering, Faculty of Basic and Applied Sciences,

International Islamic University, Islamabad

2013

i

Department of Computer Science & Software Engineering

International Islamic University Islamabad

Final Approval

Dated: _Aug 2013_

This is to certify that the Department of Computer Science, International Islamic University Islamabad,

accepts this dissertation submitted, in its present form as satisfying the dissertation requirements for the

degree of MS (CS).

Committee

External Examiner: Dr. Abdul Basit Siddiqui

Assistant Professor

Foundation University Institute of

Engineering & Management Sciences

Internal Examiner:

Ms. Sadia Arshid

Lecturer

Department of Computer Science & Software Engineering,


Supervisor:

Dr. Ayyaz Hussain

Assistant Professor,

Department of Computer Science & Software Engineering,


Declaration

ii

Declaration

I solemnly declare that this thesis submitted for the degree of MS Computer Science at the

International Islamic University is entirely my own work and has not been previously submitted by

me at another University for any degree. It is further certified that, to the best of my knowledge, any

ideas, techniques, quotations, or any other material included in my thesis from the work of other

people are fully acknowledged according to the standard referencing practices.

Qurrat ul Ain

639-FBAS/MSCS/F10

Abstract

iii

Abstract

Reliable user authentication has become very important with rapid advancements in networking

with increased concerns about security. Biometric systems perform recognition with the help of

specific physiological or behavioral characteristics(s) of a person. Biometrics establishes identity

on the basis of biological characters e.g., structure of your DNA, facial features, voice, gait etc,

instead of ID cards, PIN numbers, tokens, passwords, etc. UniBiometrics systems depend on the

evidence of only one source of information whereas multibiometric systems consolidate/combine

multiple sources of biometric evidences. Multibiometric systems are capable of enhancing the

matching performance as they get the evidence presented by different modalities or biological

characteristics and the use of multiple body traits improves the identification accuracy

significantly. Moreover, they are expected to increase population coverage, prevent spoofing

attacks and provide fault tolerance to biometric systems. In this thesis, we have proposed an

evolutionary approach to enhance the matching performance of our multibiometric system. The

system uses Discrete Cosine Transform at feature extraction level due to its high energy

compaction property. The existing methods of Linear Discriminant Analysis and Principal

Component Analysis are also used in parallel with DCT in order to compare the three feature

extraction methods. LDA being more efficient than PCA as LDA uses both intra-class scatter

and inter-class variation whereas PCA only deals with intra-class scatter. However, DCT offers

far less computational complexity than the other two methods. Three biometric traits i.e. face,

iris and ear have been used which are fused at feature level and genetic algorithm has been

incorporated for feature vector optimization. Classification is performed through the Bayesian

classifier. Results have been computed on the basis of Error Equal Rate (EER) values and ROC

curves which have shown that the use of Discrete Cosine Transform with genetic algorithm has

significantly improved the performance of our multibiometric system.

Keywords: Multimodal, Biometrics, Discrete Cosine Transform (DCT), Linear Discriminant

Analysis (LDA), Principal Component Analysis (PCA), feature level fusion, Genetic Algorithm

(GA), Bayesian Classifier

Acknowledgement

vi

Acknowledgement

I am very thankful to Al Mighty Allah for making me capable enough to understand and observe

the things around me. I thank my parents who have always supported me in achieving my goals.

It would not have been possible for me to complete my graduate studies, without their support,

encouragement and guidance towards selecting what would be best for me. I dedicate this thesis

and all the great things in my life to them. I would also like to thank my three sisters for their

timely help.

I would like to convey my sincere gratitude to my advisor Assistant Professor, Dr. Ayyaz

Hussain for providing his guidance in the exciting and challenging areas of image processing and

biometrics. His constant motivation has helped me towards the successful completion of my MS

Computer Science studies.

Table of Contents

vii

Table of Contents

List of Figures ix

List of Tables xi

1. Introduction 1

1.1 Biometric Systems 2

1.2 Biometric Systems Functional Processes 3

1.3 Desirable Properties of Biometric Traits 4

1.4 Biometric System Operations 5

1.5 Advantages of Biometric Systems 6

1.6 Limitations of Biometric Systems 7

1.7 Applications of Biometric Systems 8

1.8 Thesis Motivation 9

1.9 Thesis Contributions 10

1.10 Thesis Layout 10

2. Multibiometric Systems 12

2.1 Introduction to Multibiometric Systems 13

2.2 Taxonomy of Multibiometric Systems 14

2.2.1 Multi-sensor System 14

2.2.2 Multi-algorithm System 15

2.2.3 Multi-instance System 16

2.2.4 Multi-sample System 16

2.2.5 Multimodal System 16

2.3 Feature Extraction Methods 17

2.3.1 Principal Component Analysis 18

2.3.2 Linear Discriminant Analysis 19

2.3.3 Discrete Cosine Transform 21

2.4 Levels of fusion in Multibiometric Systems 23

2.4.1 Fusion before Matching 23

Table of Contents

viii

2.4.1.1 Sensor level fusion 23

2.4.1.2 Feature level fusion 24

2.4.2 Fusion after Matching 25

2.4.2.1 Score level fusion 25

2.4.2.2 Rank level fusion 26

2.4.2.3 Decision level fusion 26

Summary 28

3. Literature Review 29

3.1 Overview of the Existing methods 30

3.2 Problem Statement 39

Summary 40

4. Proposed Methodology 41

4.1 Introduction 42

4.2 Proposed flow of work 43

4.3 Image Preprocessing 44

4.4 Proposed Feature Extraction Methods 45

4.4.1 Implementation of DCT 45

4.4.2 Implementation of LDA 46

4.4.2 Implementation of PCA 46

4.5 Feature vector Normalization 47

4.6 Proposed Feature level Fusion 48

4.7 Feature vector Optimization using GA 49

4.7.1 Genetic Algorithm 49

4.8 Classification using Bayesian Classifier 51

4.8.1 Bayesian Classifier 51

Summary 51

5. Experimental Results 53

5.1 Introduction 54

Table of Contents

ix

5.2 Data Sets 54

5.2.1 AMI Ear database 54

5.2.2 Faces94 Dataset 55

5.2.3 MMU2 Iris Database 55

5.3 Performance Measures 56

5.4 Experimental Results 57

5.4.1 Results of Unimodal Biometrics 57

5.4.2 Results of Multimodal Biometrics 59

5.4.3 Results of feature extraction methods 63

5.4.4 Classification Accuracy Measurement 65

Summary 67

6. Conclusion and future work 69

6.1 Introduction 70

6.2 Conclusion 70

6.2 Future work 71

References 72

Glossary

List of Figures

ix

List of Figures

Fig. 1.1: Different Biometric Traits (a) Fingerprint, (b) Face, (c) DNA, (d) Hand veins, (e) Iris,

(f) Palmprint, (g) Typing, (h) Signature, (i) Ear, (j) Voice, and (k) Retina

Fig. 1.2: Working of a Biometric System [11]

Fig. 1.3: Biometric applications (a) Immigration and Naturalization service Passenger

accelerated service system (INSPASS), based on hand geometry for individual

recognition at US airports deployed to minimize the immigration processing time, (b)

Fingerprint-based door lock to limit access to certain premises, (c) Fingerprint-based

point of sale (POS) terminal for verifying customers before taking payment from their

credit cards in retail shops, restaurants and cafeterias, (d) Fingerprint verification

system employed for computer login (e) Hand geometry system for immigration and

security uses Express Card entry kiosks deployed on airport in Israel(f) Iris recognition

system used as border control system at London‟s Heathrow airport

Fig. 2.1: Information sources for biometric fusion [9]

Fig. 2.2: The PCA model, (a) 2D space illustrating the geometric representation of principal

eigenvectors, (b) 1D reconstruction (projection of the data) using the first principal

component [15].

Fig. 2.3: Example of PCA and LDA projections for two classes [15]

Fig. 2.4: Comparison of DCT and PCA in terms of Cumulative recognition accuracy as a

function of rank for the CIM face database. (- DCT, - - PCA) [16]

Fig. 2.5: Example of feature level fusion using face and iris biometric traits [19].

Fig. 2.6: Match score level fusion [10]

Fig. 3.1: Two different classes shown by the two Gaussian-like distributions, two samples per

class provided to the learning procedures, PCA or LDA. Classification result produced

by PCA using only one eigenvector is more than that produced by LDA. DLDA and

DPCA are the decision thresholds suggested by using nearest-neighbor classification

[3].

List of Figures

x

Fig. 3.2: Performance comparison of (a) unimodal and multimodal biometric systems, and (b)

classification methods [9]

Fig. 3.3: ROC Curve of the proposed fusion functions on validation set of the three databases (a)

BANCA (b) BSSR1 (c) PRIVATE [14].

Fig. 4.1: Flowchart of the Proposed Solution

Fig. 4.2: Preprocessing Results (a) Original Image, (b) Gray-scale Image, (c) Blurred Image

using Gaussian filter, (d) „Log‟ filter Image, and (e) Final Image from (c) and (d)

Fig. 4.3: Genetic Algorithm steps

Fig. 5.1: ROC Curve

Fig. 5.2: ROC performance curve of face, iris and ear biometric traits using (a) DCT, (b) LDA,

and (c) PCA

Fig. 5.3: ROC performance curves of „sum‟, „min‟ and „mul‟ fusion techniques using (a) DCT,

(b) LDA, (c) PCA, (d) DCT with GA, (e) LDA with GA, and (f) PCA with GA

Fig. 5.4: ROC Performance curves of Best Fusion technique in terms of EER with and without

the use of Genetic Algorithm using (a) DCT, (b) LDA, and (c) PCA

Fig. 5.5: ROC Performance curves of feature extraction methods using (a) „sum‟, (b) „min‟ and

(c) „mul‟

Fig. 5.6: Comparison of the Best fusion techniques of the three feature extraction methods

Fig. 5.7: Bar graph showing accuracy of fusion techniques with and without the use of Genetic

Algorithm for (a) sum, (b) min, and (c) mul

List of Tables

xi

List of Tables

Table 2.1: Examples of multimodal systems

Table 3.1: EER and Time Gain by using the proposed EER calculating method [14]

Table 3.2: Performance/Accuracy (%) of multimodal biometric databases calculated with

different fusion methods [15].

Table 4.1: Parameters of Genetic Algorithm

Table 5.1: Performance (EER) of three feature extraction methods for face iris and ear biometric

traits

Table 5.2: Performance (EER) of three fusion techniques for PCA, LDA and DCT

Table 5.3: Time Consumption by feature Extraction methods

Table 5.4: Accuracy of the simple multimodal systems and the GA multimodal systems

Chapter 1 Introduction

Multimodal Biometric Security using Evolutionary Computation 1




1.1 Introduction to Biometric Systems

Biometric systems perform recognition of individuals on the basis of their physical and/or

behavioral traits. Some commonly used traits are fingerprint, face, iris, retina, palm print, voice

pattern, signature, gait, etc. Most biometric systems will serve one of the two purposes:

identification or verification/authentication. Biometric systems provide several advantages over

the traditional methods. Biometric traits cannot be easily copied, shared, distributed or forged.

Biometric systems also provide the convenience in a sense that the user is no more required to

design or remember passwords. Multibiometric systems consolidate multiple sources of

biometric evidences. The integration of evidences is known as fusion. The information from

multiple sensors, multiple samples or multiple traits of an individual is consolidated by the

multibiometric system using various algorithms deployed on the same biometric trait.

The fundamental requirement for various operations using biometrics is to verify an individual‟s

identity. In order to provide the genuine individual with the desired privileges given that they are

provided at the correct time by having authenticated access, three approaches are available to

establish an individual's identity [16]. The said methods used in various real life applications for

verifying individual‟s identity include:

Something you have: Desired privileges can be accessed by the user when he/ she possess

some physical objects like, keys, identity card, smart card, etc., and these are shown to

the authorities to get access of something or to be identified.

Something you know: When the user already knows some predefined objects like

passwords and these objects are entered in order to verify the individual.

Something you are: A user can have access to a desired service with the help of

measurable biometric traits (some examples are given in figure 1.1).

The application of biometrics in the processes of human identification and/or verification has

some significant advantages over the other two methods. This is due to the fact that biometric

traits are complex enough to share or steal and at the same time it is almost impossible that these

traits cannot be forgotten or lost. Thus it can be stated that a higher security level can be

achieved using biometrics for person identification/verification. There are several biometric traits

which are used in various applications. Some examples of biometric traits are given in figure 1.1.



(a) (b) (c) (d) (e)

(f) (g) (h) (i)

(j) (k)

Fig. 1.1: Different Biometric Traits (a) Face, (b) Fingerprint, (c) DNA, (d) Hand veins, (e) Iris,

(f) Palmprint, (g) Typing, (h) Signature, (i) Ear, (j) Voice, and (k) Retina

1.2 Biometric System Functional Processes

A biometric system involves the following three main functional processes [15]:

Enrollment Process:

In enrollment process, a subject presents his/her biometric characteristics to the sensor along

with his/her non-biometric information. Non-biometric information related to subjects could be



name, social security number, driver license‟s number, etc. Biometric features extracted from the

captured sample and the non-biometric information is enrolled in the database.

Verification Process:

In a verification process, the claim of the person is checked whether he/she is the same person

that he/she claims to be. The user who wishes to be recognized provides some objects which

could be a Personal Identity Number (PIN), a username or a smartcard and also submits his

biometric information through some sensors like camera, microphone or fingerprint detector etc.

The system then compares the extracted template (from the captured sample) with the already

enrolled template linked to the claimed identity from the database (1:1 matching) and determines

whether the claim is true or false. When a user is willing to be recognized or verified, identity

verification is used and this is known as positive recognition.

Identification Process:

In an identification process, it is to be checked that that a particular person is? In this process, an

individual is recognized by first capturing and then extracting the person‟s biometric features for

finding a match with all the stored user templates in the enrollment database (1:M matching).

When a user is not willing to be recognized, identification is used and this is known as negative

recognition.

The significance of identification process lies mostly in negative recognition when the user tries

to avoid being found out who he is. Some examples of such applications are background checks,

criminal identification or preventing terrorists from entering certain areas. One of the major

contribution of biometrics is the negative recognition that cannot be provided with the existing,

traditional recognition methods such as passwords, PIN, tokens which can only work for positive

recognition.

1.3 Desirable Properties of Biometric Traits

Some desirable properties of biometric characteristics for good subject discrimination and

reliable recognition performance are described below [17]:

Universality: Every individual must have the specific biometric characteristic.



Uniqueness: The characteristics of each individual must be satisfactorily distinguishable

among the whole enrolled database of all the individuals.

Permanence: The biometric characteristics are required not to vary with time, for

example, appearance of wrinkle in overage can be a problem.

Measurability: It should be possible to acquire the characteristics without causing too

much difficulty. The captured images or data must be suitable for future processing.

From an application point of view, following properties should also be taken into account.

Performance: The required recognition accuracy in an application should be achievable

using the characteristics.

Acceptability: Acceptability refers to the willingness by the subject to present his

biometric characteristics.

Spoof Resistance: This refers to how difficult it is to use manufactured article (for

example, fake fingers) for physiological traits and impersonation (for example voice

tampering) for behavioral traits.

1.4 Biometric System Operations

There are three basic elements in a biometric system; 1) a sensor that captures the user‟s

biometrics characteristics; 2) a software or feature extractor which is used to convert this raw

data from the sensor to digital form for comparing with data already enrolled in the database, and

3) a database, which stores the processed digital biometric data.

All the biometric systems consist of four steps that are listed as under:

Capture: A sensor (for example, a camera for capturing face, iris or ear images) is deployed

at the start of the biometric system. It collects the sample(s) of biometric features like

fingerprint, voice etc of the person who wants to have access to a desired privilege.

Extraction: The samples obtained from the sensors are incorporated into the feature

extraction software and a template is generated. The unique features for each individual

extracted by the system are then converted into a digital biometric code. This code is then

stored into the database as the biometric template for that individual.



Fig. 1.2: Working of a Biometric System [18]

Comparison: When a person needs to be verified, the above processes are repeated in order

to generate a new template for that person. The generated template is then used to find a

match with all the stored user templates in the enrollment database.

Match/non-match: Finally the system outputs whether the new generated template at

runtime matches with one of the template from the database or not. In this way a

match/non-match or simply identified/unidentified is declared by the biometric system.

1.5 Advantages of Biometric Systems

Several advantages are offered by the use of biometrics for identification as well as verification

purposes. Here we have the person as the key so the problem of remembering password is

eliminated. Each body part of every individual is unique among the whole population and this

unique identity is employed in the biometric system to activate/deactivate or lock/unlock

something.

− Increased security – Provide a convenient way of implementing security.



− Decrease hoax attacks by using different methods and materials in construction of the

biometric system. For example reduction in ID fraud, buddy punching, etc.

− Problems of lost IDs and forgotten passwords can be eliminated by using physiological

biometric traits. For e.g. preventing unauthorized use of ID cards.

− No need to administer passwords, so reducing password management costs.

− Passwords are no more required to be remembered which are at a threat of sharing or

observance.

− With the use of biometrics it has become possible, to automatically determine WHO has

done something, WHAT has been done, WHEN something has been done and WHERE

something did happen!

− Biometrics employed in attendance system offer considerable cost and time savings.

− An individual can be directly connected to a transaction without requiring another

individual for his/her verification.

1.6 Limitations of Biometric Systems

Biometric security systems actually have an individual‟s biometric traits in a sense they are

attacking the privacy of the individual, so therefore there are a number of concerns relating to it.

Among the population, it is not necessary that all the people are willing to register for some

application on the basis of their biometric traits, for example one may not feel comfortable

giving their DNA or personal information to the biometric system stored in a database beyond

his/her control. Moreover a high level security must be employed in the biometric system

because if personal information is stolen, it may results in devastating effects on the particular

individual‟s life. Also if the biometric system has stopped working suddenly, it cannot

recognize/identify the individual. Furthermore the identity of the person may be completely lost.

So biometric systems are still having the following challenges to deal with:

− Noise in sensed data causes accuracy reduction

− Non-universality increases failure to enroll error (FTE)



− Lack of Individuality/Uniqueness increases False Accept Rate (FAR)

− Intra Class variation increases the False Reject Rate (FRR)

− Inter Class similarity increases FAR

− Openness to circumvention by spoofed attacks commonly for voice and signature

1.7 Application of Biometric Systems

Because of reliable identification and verification provided by the use of biometric systems, they

have been installed in a variety of applications. Some of them are listed as under [19]:

− Biometric systems are used widely in commercial applications in the areas such as

Internet access, computer network login, e-commerce, physical access control, ATMs or

credit cards, mobile phones, managing medical records, Personal Digital Assistant

(PDA), distance learning, etc.

− Government applications use biometric systems for making national ID cards, driving

license, passports, social security numbers (SSN), welfare expenditure codes, border

passage system, etc.

− Forensic applications make use of biometric systems for corpse identification, parenthood

determination, criminal investigation, terrorist identification, etc.

(a) (b) (c) (d)



(e) (f)

Fig. 1.3: Biometric applications (a) Immigration and Naturalization service Passenger

accelerated service system (INSPASS), based on hand geometry for individual recognition at US

airports deployed to minimize the immigration processing time, (b) Fingerprint-based door lock

to limit access to certain premises, (c) Fingerprint-based point of sale (POS) terminal for

verifying customers before taking payment from their credit cards in retail shops, restaurants and

cafeterias, (d) Fingerprint verification system employed for computer login (e) Hand geometry

system for immigration and security uses Express Card entry kiosks deployed on airport in

Israel(f) Iris recognition system used as border control system at London‟s Heathrow airport.

1.8 Thesis Motivation

Reliable identity establishment/conformance is becoming critical in a variety of applications.

Examples of those applications may include computer resources sharing connected via a

network, performing remote financial transactions, border security control, and forensic

applications. Traditional methods of establishing identity are either knowledge based systems

(e.g., passwords) or possession based systems (e.g., ID cards). Individuals have certain distinct

physiological and behavioral traits that are used by biometric systems for reliable authentication.

Biometric systems provide better security and greater convenience than the traditional systems.

Our motivation for working on this project comes from the fact that in near future biometric

systems will be supplementing or replacing the traditional systems in many applications. Most of

the biometric systems presently being used are unimodal biometric systems typically making use

of a single biometric trait for recognition purpose. There are several limitations of such systems



and some of these can be resolved by employing multibiometric systems that

consolidate/combine numerous sources of information based on biometric traits. Multibiometric

systems offer the advantage of improving the matching accuracy of the biometric system [19].

They also address challenges such as non-universality, noise, susceptibility to spoof attacks and

large intra-class variations.

1.9 Thesis Contribution

In this thesis, we addressed three issues in the design of multibiometric systems.

− Three feature extraction methods are used to compute the feature vector of each

biometric trait (face, iris, and ear). DCT, LDA and PCA methods are used for feature

extraction. The energy compaction property of DCT makes it the efficient method among

the existing approaches of PCA and LDA. At this level, the accuracy of the unimodal

systems is computed in terms of EER.

− Three fusion methods are implemented in order to fuse the feature vectors for making

multimodal biometric system. „sum‟, „min‟ and „mul‟ fusion methods are used for fusion.

It has been seen that we get different results for same feature vectors fused by using

different fusion methods. Here the EER values and the ROC curves are computed to

check the efficiency of multimodal biometric systems.

− For achieving desired level of efficiency of our multimodal biometric system, the

proposed method then implements genetic algorithm for feature vector optimization. The

use of GA improves the performance of each feature extraction method.

Experiments have been performed for calculating the accuracy and efficiency of the proposed

method. Matlab R2012a, on windows 7, has been used to implement this project. Finally the

accuracy is computed for the three feature extraction methods with and without the

implementation of genetic algorithm.

1.10 Thesis Layout

Chapter 2 gives the detailed description of the Multibiometric Systems. The advantages of

multibiometrics over UniBiometrics have been discussed. The different feature extraction



techniques used in biometrics are described like PCA, LDA and DCT. Also the various fusion

level methods are discussed in detail.

Chapter 3 deals with the literature review. Recent techniques in UniBiometrics and

multibiometric systems have been studied and analyzed.

Chapter 4 describes in detail the proposed solution for making a more accurate multibiometric

system with better recognition ability.

Chapter 5 describes the experiments performed for finding the accuracy of UniBiometrics

system and multimodal biometric system. ROC curves and EER values validate the experiments

performed with and without the use of genetic algorithm.

Chapter 6 finishes the thesis and gives some future implementation of our proposed method in

the field of multibiometrics for individual‟s identification.

Chapter 2 Multimodal Biometric Systems





2.1 Introduction to Multibiometric Systems

Those systems which aim at determining the identity of an individual by combining facts of

biometric information/traits from multiple sources are known as multibiometric systems [20].

The use of multibiometric systems for identification or verification can eliminate most of the

limitations caused by UniBiometrics systems. This is due to the fact that different biometric

sources generally reimburse for the inherited problems of the other biometric traits [21].

Therefore, a multibiometric system has several advantages over UniBiometrics systems. The

problems of UniBiometrics systems can be efficiently dealt with the implementation of

multibiometric systems as described below:

− With the help of an effective fusion method applied for consolidating different biometric

information/traits can result in significantly enhancing the recognition effectiveness of

the multibiometric system.

− The use of multibiometric systems can efficiently reduce the errors of FAR and FRR by

addressing the non-universality problem. For example, due to an accidental cut on finger

print ridges, it would be difficult to enroll a person in a finger print system. However, if

multibiometric system is used, other biometric traits like face, voice, ear, iris, etc. can be

employed to identify the individual.

− Multibiometric systems offer a flexible environment for user identification/verification.

For example, if a user registers into the system using various biometric traits, it is

possible using a multibiometric system that at the runtime only a few of the traits can be

used for matching for checking the authenticity of the user. Thus multibiometric system

provides a convenient way of authentication both for the system and the user.

− Employing multiple sources of information greatly reduce the bad effects of noise in raw

data. While capturing a particular biometric information/trait, if the sample is not of

sufficient quality, the samples of other sources may still present enough distinguishing

information in order to identify that individual.

− As it is not easy to hack/spoof multiple biometric templates at the same time so

multibiometric systems offer resistance against spoof attacks more than the

UniBiometrics systems. Moreover, a user challenge response technique can be deployed

by a multibiometric system at runtime during capturing of the biometric trait. This can be



implemented by taking only some of the traits in a random order so that it can be ensured

that a live user is interacting with the system.

− Multibiometric systems provide template security by consolidating the biometric

information from various sources using some fusion methods.

Multibiometric systems also have some of the drawbacks in comparison with UniBiometrics

systems.

− As they use multiple sources of information so for capturing different biometrics,

different instruments are required making the system more expensive.

− More storage space and resources for computation are required by a multibiometric

system as compared to UniBiometrics systems.

− Multibiometric systems need more user time during enrollment at image acquisition stage

causing difficulty for the user.

− For combining the information obtained from multiple sources, an appropriate technique

must be followed as different fusion techniques cannot result in providing equal accuracy

of a multibiometric system. Different subsets of biometric traits are fused with different

fusion methods, one fusion method cannot be applied to all the subsets of biometric traits,

so the choice of fusion method is a crucial step.

2.2 Nomenclature of Multibiometric Systems

A multibiometric system relies on the biometric information provided by multiple sources.

Depending on the nature of these sources, a multibiometric system can be categorized into the

following five categories [9] as illustrated in figure 2.4:

2.2.1 Multi-sensor systems

Multi-sensor systems exploit several sensors to get images of a single biometric trait of an

individual [9]. For example, a face detection system may use different cameras fixed at different

position from face to capture the front, left or right side image of face of an individual; a camera

with varying frequencies may be used to take images of the face, iris or ear; or an optical sensor

may be employed to acquire fingerprint image of an individual. The employment of multiple

sensors in a multibiometric system can result in achieving related information that can



significantly improve the identification prospective of the system. As an example, images of a

person captured in varying illumination, or images captured in infrared light can give different

information resulting in improvement or may be reduction in recognition accuracy. Similarly, a

2D face authentication system can be made more efficient by using a 3D camera for face images

employing the shape information of the individual‟s face.

Fig. 2.1: Information sources for biometric fusion [9]

2.2.2 Multi-algorithm systems

Sometimes more than one feature extraction method applied on the one biometric data may result

in considerable recognition accuracy of the multibiometric system. Multi-algorithm systems

employ different algorithms or feature extraction methods and each algorithm produces a feature

vector. These feature vectors are combined using particular fusion technique. They are cost-

effective in a sense that they do not require much hardware (sensors) compared to other

categories of multibiometric systems. However on the other hand, time and computational



complexity are increased due to the use of various feature extraction and identification modules

in the system. For example Sabareeswari et al [13], have proposed a multimodal biometric

system using signature, face and ear biometric traits using two feature extraction methods of

FLD (Fisher‟s Linear Discriminant) and PCA (Principal Component Analysis).

2.2.3 Multi-instance systems

A multi-instance system attempts to exploit multiple images of a single biometric trait. They are

also called multi-unit systems [9]. For example, the system captures images of both index fingers

or both irises of an individual for identity verification. These systems are more cost effective as

only one sensor can be used to take the multiple images of a single biometric trait sequentially.

As an example, Jacob et al [52] have used multi-instance, unimodal biometric system

implementing nine images of fingerprints for each individual for verification purpose.

2.2.4 Multi-sample systems

In a multi-sample system, more than one sample of the same biometric trait is taken from a

single sensor to get images/samples with variation in illumination, position and/or expression

that can occur in the trait [9]. For example, for a face identification system, the front image as

well as left and right side face images are captured in order to deal with the variation in the facial

pose and expression. Similarly, a small size fingerprint sensor is used to get various fingerprint

regions with different orientations which are then stitched together using image mosaicing

technology to get a complete fingerprint image. In a multi-sample system, a crucial step is

finding the number of samples to be used in the system. Also it is significant that the samples

obtained must present the variability among samples for one individual as well as the uniqueness

of the individual‟s biometric data [9].

2.2.5 Multimodal systems

A Multimodal system declares identity of an individual based on the factual information

provided by more than one biometric trait. For example, the old multimodal systems mostly use

face and voice biometric features for identification and/or verification purposes as listed in Table

2.1. Better results can be achieved when uncorrelated traits e.g., fingerprint and face are

employed as compared to correlated traits e.g., voice and lip movement [9]. The main

disadvantage of these systems is the high cost of setting up these systems as multiple sensors are



required for capturing multiple biometric traits. The significant advantage of these systems is that

accuracy can be drastically enhanced by using more number of traits. This is the reason we are

proposing a multimodal biometric identification system for achieving better results. The number

of traits used in a particular verification scenario can be restricted by real-life consideration such

as enrollment time of the whole population, the cost of deployment, throughput time according to

various implemented algorithms, predictable error rate, etc.

2.3 Feature Extraction Methods

Feature extraction is a method employed for the purpose of dimensionality reduction, in pattern

recognition and image processing. When the raw data entering into an algorithm is a large

vector, then for earning time computation this data will be changed into a reduced set of features,

which is also termed as feature vector. The process of transforming data into a set of features is

called feature extraction [22]. It is necessary that the extracted features are chosen carefully, so

that they may have the ability to extract the information from the input data for performing the

desired task of identification/verification. Thus it reduces the complexity of the system by using

reduced number of features instead of the full size input data.

A feature is function of measurement, that is capable of uniquely specifying some biometric

property of an object/individual, and is calculated in a way that it can measure some valuable

properties/characteristics of the individual/object [24]. Features can be classified into various

categories according to the information provided by them:

− General features: These are application independent features for example color, quality,

and form or shape. They are subdivided into the various classes; pixel-level features,

global features, and local features. Pixel-level features are obtained at pixel level, e.g.

color, location, etc. Global features are computed over the whole image or a cropped area

of the original image. Local features are obtained by subdividing the image with the help

of segmentation and/or edge detection.

− Application specific features: As the name implies, these features depend on the type of

application for which they are generated for example human face, iris, ear, and other

intangible features like voice, gait or lip movement etc [24].

http://en.wikipedia.org/wiki/Dimensionality_reduction

http://en.wikipedia.org/wiki/Pattern_recognition


http://en.wikipedia.org/wiki/Image_processing

http://en.wikipedia.org/wiki/Algorithm



A multibiometric system based on domain-specific features result in better performance as

compared to general features. However, if expert knowledge is not available general features can

be used. The examples of such general dimensionality reduction methods include Principal

component analysis, Multifactor dimensionality reduction, Multilinear subspace learning, Linear

discriminant Analysis, Latent semantic analysis, Nonlinear dimensionality reduction, Kernel

PCA, Multilinear PCA, etc.

2.3.1 Principal Component Analysis

Principal Component Analysis is a statistical method in which one-dimensional as well as multi-

dimensional data is analyzed. PCA monitors association between various dimensions of the

given data and extracts principal dimensions, where the variation among the whole data is at

peak [13]. The significance of PCA is that it determines those features that can describe most of

the variation in the data by using only a small number of features. The PCA matrix consists of

the eigen-values which are obtained from the covariance matrix S and hence it takes long time

making it computationally expensive [25].

∑ ) ) )

2.1

where n represents the number of instances, xi is the ith

instance, and m represents the mean

vector of the input data.

Following algorithm depicts the process of obtaining principal components from a given input

data:

1) First of all covariance matrix S is calculated from the given input data.

2) The eigenvectors and eigen-values of this covariance matrix S are then computed and

sorted in a descending order with respect to the eigen-values.

3) A specific number of components from the eigenvectors are selected and a new matrix is

formed.

4) Lastly, the product of the original space the obtained new matrix is computed, that results

in dimensionality reduction.

http://en.wikipedia.org/wiki/Principal_component_analysis

http://en.wikipedia.org/wiki/Principal_component_analysis

http://en.wikipedia.org/wiki/Multifactor_dimensionality_reduction

http://en.wikipedia.org/wiki/Multilinear_subspace_learning

http://en.wikipedia.org/wiki/Latent_semantic_analysis

http://en.wikipedia.org/wiki/Nonlinear_dimensionality_reduction

http://en.wikipedia.org/wiki/Kernel_PCA

http://en.wikipedia.org/wiki/Kernel_PCA

http://en.wikipedia.org/wiki/Multilinear_PCA

http://en.wikipedia.org/wiki/Autoencoder



For defining a threshold, the principal axis describes the required collective percentage of

variance. This percentage defines the total number of components, chosen for making the new

matrix.

(a) (b)

Fig. 2.2: The PCA model, (a) 2D space illustrating the geometric representation of principal

eigenvectors, (b) 1D reconstruction (projection of the data) using the first principal component

[15].

2.3.2 Linear Discriminant Analysis

Linear discriminant analysis (LDA) is also a statistical method like PCA and is most commonly

employed in the fields of pattern recognition, image processing and machine learning to find a

combination of features that are capable of distinguishing two or more classes of objects. The

resultant set of features can be deployed as a linear classifier in an application. A LDA classifier

is most commonly used as a dimensionality reduction method as well as a classification method.

LDA forms a subset of dependent variables representing a set of other features. LDA and PCA

are similar in the sense that both of them search for a set of features that has the ability to explain

the whole data significantly. LDA can clearly create the difference between the defined classes.

On the other hand PCA cannot explain difference in classes.

During image acquisition, much of the variation is data is produced due to the changing

illumination, position and expressions of the individual. In these conditions PCA cannot give


http://en.wikipedia.org/wiki/Machine_learning

http://en.wikipedia.org/wiki/Linear_combination

http://en.wikipedia.org/wiki/Features_%28pattern_recognition%29

http://en.wikipedia.org/wiki/Dimensionality_reduction

http://en.wikipedia.org/wiki/Statistical_classification

http://en.wikipedia.org/wiki/Dependent_variable



highly consistent results [13]. To overcome this problem, LDA may be used to produce a

subspace projection matrix. The LDA method takes advantage by utilizing within-class

information and in this way it can significantly reduce variation within each class. However, it

may still maximize class separation. The following equations show the scatter matrices, which

shows the within class (SW), between-class (SB), and total distributions (ST).

∑ ∑ ) )

(2.2)

∑| | ) )

(2.3)

∑ ) )

(2.4)

here the average of each individual class Xi is

| |)∑

and the average image vector

of the complete training set is (

)∑

[13].

Fig. 2.3: An example of LDA and PCA projections for two classes [15]



2.3.3 Discrete Cosine Transform

Coding transformations are based on the fact that a pixels in an image must have a relationship

with its adjacent pixels. As an example of the just stated fact, we have seen that in a video

transmission, the adjacent pixels in successive frames show a high degree of association with

each other. Therefore, this king of association between pixels can be utilized to predict a pixel

from its adjacent pixels. Mapping of this correlated data into uncorrelated coefficients is called

transformation. This transformation must be capable of exploiting the fact that an individual

pixel contains relatively small information i.e., contribution of a pixel in an image or a sample

can be predictable using its adjacent pixels [23]. Transform coding is an essential element of

present-day image/video processing applications.

Discrete Cosine Transform (DCT) is one of the coding transformation method that has emerged

as the tool for image transformation in visual applications in the last decade. It has been

extensively used by video coding standards, i.e., JVT, MPEG, etc.

The Discrete Cosine Transform (DCT) attempts to transform the input image data into a form

where coefficients have no correlation. Thus each of the resultant transform coefficients can be

programmed separately without the loss of accuracy to be achieved during compression. This

section illustrates the concept of DCT computation for image processing in detail and its

important properties [23].

For computing DCT, the equation below shows vs (k), which are DCT coefficients and the

sequence u (n) as a vector of an image. The output v(k) can be obtained by applying DCT as a

transformation matrix to this image vector [20]. The DCT transformation matrix, C = {c (k, n)},

is defined as:

) {

√

√

(

)

)

2.5

here k and n represent the row and column indices, respectively. Using above equation, the DCT

of the sequence u (n) can be computed as:

v = Cu 2.6



In order to get the original image back we can reverse the process by using inverse discrete

cosine transform. It is employed to obtain u (n) from v (k) as defined by:

) ∑ ) )

)

)

2.7

By arranging Eq. 2.6, the inverse discrete cosine transform, u, of a vector v is obtained by taking

the inverse of matrix C. Mathematically, the inverse discrete cosine transform is obtained from

the following equation:

u = C -1

v 2.8

The above definitions describe that by applying the discrete cosine transform to an input image

vector, we can decompose it into a summation of basic cosine series [20]. Eq. 2.7 shows that u(n)

can be constructed again by adding the cosine sequences. The summation of cosine sequences is

weighted by the DCT coefficients. These basic sequences of the DCT are the rows of the matrix

C.

In the paper [20], two methods of feature extraction have been used, i.e., PCA and DCT.

Comparison is performed on the basis of number of coefficients used and the size of database.

They have shown that PCA takes more computation time for processing coefficients during the

enrollment stage as compared to DCT. However, once the enrolled database has been computed,

PCA uses very less time during the matching process of input image coefficients with the

enrolled database as compared to DCT. The cumulative recognition accuracy has been

computed which shows that the DCT outperforms the PCA method as shown in Fig 2.3.



Fig. 2.4: Comparison of DCT and PCA in terms of Cumulative recognition accuracy as a

function of rank for the CIM face database. (- DCT, - - PCA) [20]

2.4 Levels of Fusion in Multibiometrics

Before implementing the fusion process, it is essential to find out the type of information that

needs to be combined/ consolidated. The level of information decreases after each step of

processing in different element of a multibiometric system [19]. The richest source of

information is contained in the raw data captured form the sensors whereas the final decision

provided by the biometric system just contains an abstract level of information.

There are two basic levels of fusion i.e., pre-classification (fusion before matching) and post-

classification (fusion after matching). These fusion levels are further subdivided. Fusion before

matching has two subdivisions i.e., sensor level fusion and feature level fusion. Fusion after

matching is subdivided into score level fusion, rank level fusion and decision level fusion [19]:

2.4.1 Fusion before Matching

2.4.1.1 Sensor Level Fusion

The raw data that is just captured by one or more sensors is consolidated. This is called sensor

level fusion. This level of fusion cannot be performed on two different biometric traits. It is the

limitation of this fusion that it can only be used when we have multiple instances of the same



trait captured from single sensor, or multiple samples of the same trait captured from multiple

sensors. For example, various 2D face images acquired from multiple sensors positioned at left,

right and front of the face can be stitched to make a 3D image of the individual‟s face [21]. It is

an essential point in sensor level fusion that various samples must be well-suited and the

association between points in the original samples must be known before applying fusion.

2.4.1.2 Feature Level Fusion

Feature vectors created by various information sources are integrated in feature level fusion into

a new feature vector. For homogeneous feature sets (multi-instance systems), for example,

multiple images of a person‟s hand geometry, first the weighted average of the individual feature

vectors is computed and then the resulting feature vectors are fused [27]. For non-homogeneous

feature sets, (multi-modal system), for example, features of different biometric traits like face,

iris, and ear, a single feature set can be obtained by concatenation of the biometric feature vector.

Dimensionality reduction scheme like feature selection/transformation is applied to obtain a

minimal feature set. The key advantage of the feature level fusion is that it facilitates the removal

of correlated feature values improving recognition accuracy. Feature level fusion is difficult to be

employed for the following causes [28]:

1) The feature vectors obtained for multiple biometric traits might be incompatible for

fusion. For example, the eigenvectors of face and the minutiae set of fingerprints.

2) By concatenating two or more feature vectors, the resultant vector might have very large

dimensions and this might cause dimensionality problem. In such cases, when

sufficiently large numbers of training samples are not available, increasing number of

features result in performance degradation.

3) Access to the feature vector might not be provided in most cases by commercial

biometric system vendors.

4) More complex matching algorithms might be required to work on concatenated feature

vectors.



Fig. 2.5: Example of feature level fusion using face and iris biometric traits [19].

2.4.2 Fusion after Matching

2.4.2.1 Score Level Fusion

Different biometric matchers provide match scores representing the relationship between the

input and the stored template vectors from the database. These match scores are consolidated to

reach the final recognition decision. After the sensor level and feature level information, highest

level of information is contained by the match score about the input biometric sample. Fusion at

this level provides the best swap over between the available information content and

convenience of fusion. Therefore, this scheme is extensively studied in literature. This level of

fusion is also called measurement level fusion or confidence level fusion.

Fusion methods at score level have been further subdivided into three kinds [19]: transformation,

density and classifier-based techniques.



Fig. 2.6: Match score level fusion [19]

2.4.2.2 Rank level Fusion

A rank is associated in the rank level fusion with each biometric trait at every module of the

biometric system. The higher rank assigned to a biometric trait indicates a better match [19].

Rank level fusion requires combining the multiple ranks of individual biometric subsystems to

determine a new rank for each biometric trait. The decision is made according to the new ranks

for all identities. Ranks reveal more information for determining the best match and provide less

information as compared to the match scores.

There are three methods usually used to consolidate the ranks allocated by each biometric

matcher. Those are the logistic regression, the borda count, and the highest rank method. Logistic

regression is the most efficient among the three methods.

2.4.2.3 Decision Level Fusion

Decision level fusion is employed when the final recognition decision by each of the individual

biometric recognition system is available at hand. Most of the commercial off-the-shelf (COTS)

biometric systems use this level of fusion. They are designed in such a way that the user can see

only the final decision of the individual biometric matchers. In this case, only decision level

fusion can be employed feasibly. Decision level fusion can be incorporated using the methods

like majority voting, „and‟ and „or‟ rules, Bayesian decision fusion, weighted majority voting,

the Dempster-Shafer theory of evidence, etc [21].



Table 2.1: Examples of multimodal systems

Modalities Authors Level of Fusion Fusion Methodology

Face and voice [33] Match score; HyperBF; Rank Geometric

weighted average

[35] Match score Min, Max, Sum, Product and

Median rules

[29] Match score

SVM; multilayer perceptron;

C4.5 decision tree; FLD;

Bayesian classifier

[31] Match score Bayesian theory

Face, voice and lip

movement

[2] Match score

Weighted sum rule at Decision

level fusion; majority voting

Face and fingerprint

[38] Match score Product rule

[32] Match score Sum rule, Weighted sum rule

Face, fingerprint and

hand geometry

[47] Match score Decision trees; Sum rule; linear

discriminant function

Face, fingerprint and

voice

[34] Match score Likelihood ratio

Face and iris [37] Match score

Sum rule; weighted sum rule;

FLD; neural network

Face and gait [46] Match score Sum rule

[48] Match score Sum and product rules

Face and ear [39] Sensor Concatenation of raw images

Face and Palm print [41] Feature Feature Concatenation

Fingerprint, hand

geometry and voice

[42] Match score Weighted sum rule

Fingerprint and hand

geometry

[30] Match score Reduced multivariate

polynomial model

Fingerprint and

Voice

[43] Match score Functional link network

Fingerprint and

signature

[45] Match score SVM (quality measures

incorporated)

Voice and signature [36] Match score Weighted sum rule



Summary

The design of multibiometric system usually depends on various aspects such as capturing and

processing of input samples, sources of information, and fusion level scheme. In the last decade,

a lot of work has been done by using different taxonomic architectures of multibiometric systems

and the efficiency of a particular method depends on the specific application for which the

biometric system has been deployed. The chapter discusses various feature extraction methods

and different levels of fusion employed by multibiometric system in order to efficiently

consolidate multiple biometric feature vectors in a way of providing minimum loss of

information content. Table 2.1 summarizes some of the work in the field of multibiometrics

depending on the sources of information used.

From these tables, it is obvious that the match score level fusion has been studied and practiced

extensively by the biometrics experts. However, informal methods of normalization have been

employed by these match score level fusion strategies for finding weights of each biometric trait

used. Also the feature level fusion has not been broadly studied. Hence, our thesis aims at

developing a structure for feature level fusion in multibiometric systems.

Chapter 3 Literature Review





3.1 Overview of the Existing Methods

Hong et al [1] described a bimodal biometric system using fingerprint and face biometric traits.

The system utilized the minutiae-based fingerprint and PCA based face recognition system.

Fusion at the decision level has been used. PCA is deployed in those systems where a smaller

number of artificial variables are to be formed from the observed variables obtained during

image analysis. The artificial variables are called principal components of the original image.

These components are then used for prediction of other variables in the next stages of the

identification system. The authors have compared the unimodal systems for their biometric traits

with the multimodal biometric system obtained after fusion. With 0.01% FAR, the unimodal face

systems attained FRR of 61.2% and for fingerprint unimodal system the achieved FRR is 10.6%.

However the multimodal approach outperformed the unimodal approach by resulting in FRR of

6.6% for the same value of FAR.

Limitation of the system is that decision level fusion assumes that the matching values of faces

are numerically independent and have no correspondence with the matching values of

fingerprints. While the assumption is valid for fingerprints and faces, it may not be true for other

biometric characteristics.

Frischholz et al [2] proposed a multimodal approach for commercial purposes named BioID. The

system uses three biometric traits i.e., face, voice and lip movement for individual verification.

Face images and lip movement were captured during a video progression and the voice is taken

from an audio device. According to the security level, experiments on 150 persons produced

good results with an FAR value below 1%.

The classification process implemented by the authors can lead to insecurity which is the major

drawback of BioID. After preprocessing of a runtime pattern of the three biometric templates

ranked scalar products are obtained and the highest rank suggests the resulting class. This policy

is termed as winner-takes-all. As this policy always results in a classification because it never

rejects a pattern so the authors accounted for the second highest scalar product. They provided a

rule that if the difference between the highest and the second highest ranked output is smaller

than a specified threshold, the pattern will be rejected otherwise accepted. Judgment on the



classification result is based on the fact that if the difference between the two highest scalar

products is approaching to zero, the two people are impossible to distinguish, and the therefore

this proposed method of classification leads to “insecurity.”

Object recognition systems widely use appearance-based methods [3]. PCA and LDA are the

most implemented appearance based methods and have been confirmed to be useful for many

applications for example face recognition. It is generally taken that LDA would always

performed better than PCA, however there are systems that suggest otherwise based on the

application scenario.

Fig. 3.1: Two different classes shown by the two Gaussian-like distributions, two samples per

class provided to the learning procedures, PCA or LDA. Classification result produced by PCA

using only one eigenvector is more than that produced by LDA. DLDA and DPCA are the

decision thresholds suggested by using nearest-neighbor classification [3].

The figure above shows that the PCA might outperform LDA when we are dealing with small

number of samples per class or secondly, when the training data non-uniformly samples the

fundamental distribution. In many of the practical applications, and especially face recognition,

we cannot predict/know beforehand the underlying distributions for the various classes. So, for

practical applications it would be difficult to determine whether the available training data is

sufficient for the job or not.

Fierrez-Aguilar et al [4] suggested a multimodal approach using three biometric trait i.e., face,

fingerprint, signature. The face system is based on selection of global features and minutiae are



extracted for the fingerprint system. The signature system uses HMM modeling of sequential

functions combined together with the help of fusion methods. Two score level fusion methods

have been used i.e., support vector machine (SVM) and the sum-rule which are implemented

both as user-free and user-reliant methods. The EERs of the unimodal systems for face, the

online signature, and the finger print verification systems were 10%, 4%, and 3%, respectively.

However the multimodal system fused with the sum-rule, the SVM user-free, and the SVM user-

reliant approaches achieved EERs of 0.5%, 0.3%, and 0.05%, respectively. The stated results

showed that the multimodal system has outperformed all the unimodal biometric systems.

Disadvantage of the paper is that these values are calculated by using a database of only 50

individuals which cannot prove good for real-life applications comprising of huge databases.

Kumar et al [5] developed a bimodal approach using two biometric traits i.e., palm print and

hand geometry. These biometric traits are fused at two fusion levels; at the feature level fusion

by concatenation of the feature vectors as well as at the match score level fusion by using max

fusion rule. Only the fusion approach at the matching score level outperforms the unimodal

systems. The palm print performs better than hand geometry unimodal system. The palm print

the best among the two unimodal approaches gained a FAR of 4.49% at an FRR of 2.04%.

However the fusion of these unimodal systems resulting in a multimodal approach performs

better by achieving a FAR of 0% at a FRR of 1.41%.

The advantage of the paper is that only one sensor is used to get the images of both biometric

traits hence making the proposed method user-friendly as the user does not face the

inconvenience of going through multiple sensors for capturing multiple images for other

multibiometric systems. The authors have stated the drawback because of user-involvement at

the image acquisition stage that some users could not touch their palm/fingers on the imaging

board properly causing FTE (Failure-to-Enroll) error. These images were removed. Out of 500,

28 such images were identified and discarded. This causes performance degradation.

In the same year Heseltine et al [6] investigated the three face recognition methods; the eigen-

face method, the direct correlation method, and fisher-face method or Fisher‟s Linear



Discriminant (FLD), when applied during the pre-processing step. The optimum image pre-

processing techniques were used which showed that the FLD method has the lowest EER

(17.8%), performing better than the other two methods. FLD consumes less time in computation

however it was slightly better than direct correlation with an EER 18.0%. The Eigen-face method

showed the least accuracy with an EER 20.4%.

The advantage of the paper lies in the use of FLD which uses its property of utilizing within class

information, as it reduces variation within each class and at the same time maximizing class

separation. In other words, it tries to reduce FAR. The limitation of the paper is that the authors

were still unable to identify which preprocessing methods have enhanced which specific features

and when a given preprocessing method performs most efficiently.

Toh et al. [7] proposed a multimodal biometric system based on three biometric traits; hand

geometry, fingerprint, and voice. These traits have been fused at the match score level with

weighted sum rule. They took the multimodal decision problem at match score fusion as a two-

stage problem i.e., learning and decision. Experiments performed on the unimodal biometric

systems of fingerprint, speech, and hand-geometry confirmed that only local learning can

enhance verification ERRs to almost 50%.

In 2005, Snelick et al. [8] developed a bimodal multibiometric system based on fingerprint and

face biometric traits. These systems are fused at match score level. Three fingerprint recognition

systems and one face system were used for implementing the proposed the method. The EERs of

the best fingerprint system and the face recognition system were 2.16% and 3.76%, respectively,

while the multimodal system developed through the max fusion rule on normalized scores

attained an EER of 0.63%.

Their work has shown that Counter-of-the-shelf (COTS) multimodal biometric system using

fingerprint and face biometric traits outperform the unimodal COTS systems. However, the

accuracy of the proposed COTS multimodal system is not better than the non-COTS existing

multimodal systems developed by others.



Ross et al [9] described the methods for efficiently consolidating different sources of biometric

information. They have proposed that early fusion methods are expected to perform better than

delayed fusion methods or simply feature level fusion can perform better than score-level and

decision level fusions. However, prior to implementation of fusion methods it is not possible to

estimate the performance gain achieved through each of the methods. They have compared

unimodal biometric systems of face and voice modalities with their multimodal counterparts.

And results have shown the improved performance with multimodal biometric system. The

figure 3.2 (a) shows the ROC curves of the unimodal and multimodal systems fused through

Bayesian fusion technique. Part (b) shows the performance of different classifiers for the

multimodal system. Roc curves show that Bayesian classifier has outperformed the other

methods of SVM Poly, SVM Gaussian, Fisher Linear Discriminant and Multilinear Principal

Component Analysis.

(a) (b)

Fig. 3.2: Performance comparison of (a) unimodal and multimodal biometric systems, and (b)

classification methods [9]

A new unimodal approach for efficient individual identification has been introduced in 2009 by

Kumar et al [10] using Knuckle Codes. The preprocessed knuckle images are employed to

generate Knuckle Codes with the help of localized Radon transform that clearly show random

curved lines and wrinkles. The similarity between two Knuckle-Codes is calculated from the



least matching distance which results due to noisy data like variations resulting from positioning

of fingers. The practicability of the proposed approach is checked on a database of finger

knuckles from 158 subjects. Experiments showed an EER of 1.08% and rank one recognition

rate of 98.6%, therefore proved that the proposed method can be used for human identification.

Various fusion methods can be implemented on UniBiometrics systems for generating

multimodal biometric systems. The fusion can be applied by using different algorithms on the

same biometric trait. For example, Hocquet et al in 2007 [11] worked with three different

keystroke which are fused together in order to get a decrease in error equal rate (EER). The

limitation here is that less than 40 subjects have been used in the database. In the same year, Teh

et al [12], developed a fusion approach with two keystroke systems. These systems are fused on

the whole by using weighted sum rule. However, information about assigned weights and their

computation is not listed. Also the system is implemented by using only 50 subjects. They also

proposed that fusion methods can be implemented on different biometric traits to improve the

biometric verification.

In 2010 Sabareeswari et al [13] proposed a multimodal biometric system using three biometric

traits i.e., face, ear and signature. The proposed system utilizes two techniques at the feature

extraction level; PCA and FLD (Fisher‟s Linear Discriminant) for identity authentication. The

proposed system implemented the novel rank level fusion method for consolidating the three

different biometric matchers. Three methods have been used for combining the ranks of

individual matching biometric systems. They are the logistic regression, the highest rank and the

Borda count. Results have shown the better performance of logistic regression among the three

rank level approaches.

The advantage of the method is that even with low quality image data for the face, signature and

ear traits, fusion of three unimodal systems enhances the overall performance of the

multibiometric identification system. One of the limitations of the paper is that the use of

signature verification is a time-consuming process for the user in real time situations. For the

purpose of person identification, selection of signature biometric identifier is inappropriate as

signatures are used normally for verification purposes. Another limitation is that the authors have



incorporated PCA and FLD based global features only, however if local features were extracted

from the images, performance could be enhanced.

In the same year Giot et al [14] have presented a multimodal biometric system that focused on

two important considerations in the field of multibiometrics. They have proposed a high-speed

EER calculating method and two fusion techniques which are optimized through genetic

algorithm. Five different biometric systems have been used to test the developed EER calculating

method and a comparison is made with the existing methods. Three multibiometrics (two real

and one virtual) databases have been deployed. Fusion has been performed at the match score

level and has been validated on the databases. Their method is superior to other methods because

it uses less time and give accurate results in terms of EER. These results have proved that fewer

impostors can be accepted as well as fewer authentic users can be rejected by using this system.

Furthermore better security can be achieved for the verification process. Best gain calculated on

three databases is 78% whereas the least is 28%.

(a) (b)



(c)

Fig. 3.3: ROC Curve of proposed fusion functions on the three databases (a) BANCA (b) BSSR1

(c) PRIVATE [14].

The advantage of the proposed method is to reduce the computation time of the genetic

algorithm because its fitness function is calculating the EER. Results have shown that the fusion

functions along with the use of genetic algorithm have outperformed the existing fusion

functions i.e., sum, min and product. Limitation of the paper is that the ROC curve is not precise

due to computing time earning. Table 3.1 gives the comparison of EER on three different

datasets.

Table 3.1: EER and Time Gain by using the proposed EER calculating method [14]



Jacob et al [52] proposed a unimodal multi-instance approach for finger prints implemented by

feature level fusion. The proposed system aims that the feature level has better performance than

the matching score on the basis of their processing time. Nine randomly selected multi-instance

images undergo thinning process twice, once using Hilditch‟s algorithm and second time using

Hit and Miss Algorithm. Correlation based feature extraction is used and for finger matching,

cross correlation motivated by square Euclidean distance has been employed. Final decision is

made by comparing the fusion results with a set threshold.

The above systems proved that feature level fusion proved to be much effective in terms of

availability of raw materials and performance if multiple traits of the same biometric are used.

Secondly, the processing time has been reduced as normalization for the same biometric feature

vectors is not required as they are already compatible for fusion purpose. There are several

limitations of the proposed method. If fingerprint orientations are a little misplaced the system

cannot generate accurate results i.e., it has the problem of lack of uniqueness/individuality. Also

the cross correlation technique sometimes caused problems in matching like nonlinear

distortions, variation in skin condition and finger pressure.

In the thesis [15], a new multimodal recognition system is proposed which combines the

evidences from face and fingerprint biometric traits. Fusion of these traits is made at the decision

level. In this approach, the two biometric classifiers are independently so that class specific

information can be accessed for enhancing performance. Results have shown that the proposed

methodology outperforms the existing state-of-art fusion techniques.

LDA and Non-Parametric LDA have been employed for providing class specific information.

Face biometric classifier is good at providing this background information. Due to the presence

of high sensitivity of minutiae points, fingerprint biometric classifier cannot provide this

information hence cannot identify and individual‟s identity properly.

Fusion is applied on the two biometric classifiers using sum rule. The authors have also proposed

a method of Multiple Classifier Combination (MCC). Experiments have proved that the

proposed method (LDA and Nonparametric LDA) outperforms the on hand fusion techniques

like sum and product rules, Dempster-Shafer theory and decision template. Their results are

presented in table 3.2 below.



Table 3.2: Performance/Accuracy (%) of multimodal biometric databases calculated with

different fusion methods [15].

3.2 Problem Statement

Sabareeswari et al [13] proposed a multimodal biometric system using three biometric traits i.e.,

face, ear and signature that have been fused at the rank level. The proposed system utilizes two

techniques at the feature extraction level; PCA and FLD (Fisher‟s Linear Discriminant) for

identity authentication. It has been observed that the use of signature verification is a time-

consuming process for the user in real time situations. For example if a person is to be identified,

it should be the responsibility of the organization to identify that person without user

involvement in order to speed up the identification process. In other words, user-involvement

reduces the time efficiency of the identification process and causes delay

Secondly, the authors have implemented PCA and LDA based global features only. They have

shown the superiority of FLD over PCA that has already been proved. Furthermore, it has been

observed that the existing methods of PCA and LDA do not perform equally well in case of

different fusion techniques and with different datasets. So a new approach with different

biometric traits (fusion of which have not yet been studied) is required to overcome the problems

faced by PCA and LDA systems.



Summary

Although rapid progress has been observed in the development and deployment of biometric

systems for identification/verification and authentication purposes in the past few decades, a

number of research issues in biometrics still require a lot of expert‟s attention. The error rate of

the biometric systems cannot have zero value due to the factors of intra-class variations and

inter-class similarity. Furthermore, the high failure rates (FTER and FTCR) also limit the use of

biometric systems in various applications.

Solutions for approaching the error rates to zero in biometrics include the development of new

sensors that can reliably, conveniently and securely capture the biometric traits of an individual.

Also biometric system performance can be increased by the development of efficient matching

algorithms, selecting best fusion strategies for consolidating evidence from multiple biometric

sources in order to reduce the limitations caused by using individual sources and the

development of methods for template security of the multibiometric systems. In this thesis, we

focus on biometric systems that integrate evidence obtained from multiple biometric traits for

efficient human identification. Multibiometric systems provide many advantages that can reduce

the problems caused by the commonly-used UniBiometrics systems. This thesis mainly deals

with two crucial matters in the design of a multibiometric system, namely, feature extraction

methodology, and feature vector optimization.

Chapter 4 Proposed Methodology





4.1 Introduction

A new multimodal biometric system is proposed that reduces user involvement at the

identification stage to make the process of identification as fast as possible in everyday life. For

that reason, the traits of signature, palm print, finger prints, keystrokes and voice cannot be

utilized as they take a lot of user‟s time in scanning of these traits. Instead iris, face and ear

images are to be incorporated in the system so that computation time can be minimized.

Secondly, Discrete Cosine Transform (DCT) is employed along with the old techniques of

feature extraction like PCA and FLD. Genetic Algorithm (GA) is used for feature vector

optimization after feature fusion from the three traits (face, ear, and iris). Selected features are

then incorporated into a classifier that classifies and shows the result. The classifier confirms the

identity of the person and outputs as “identified” or “unidentified”.



4.2 Proposed flow of work

Fig. 4.1: Flowchart of the Proposed Solution

The flowchart of the proposed work has been shown in figure 4.1 above. According to the figure,

images of the three databases (face, iris and ear) are sent to the preprocessing step. Due to the

unavailability of the biometric traits for a single person, we have implemented a virtual database

by pairing the three sets of face, iris and ear datasets.

Face Image

Face, iris and Ear Features

Code


Code

Face Preprocessing Ear Preprocessing Iris Preprocessing


DCT System FLD System PCA System

Feature level Fusion

Feature Vector Optimization

using Genetic Algorithm

Bayesian Classifier

Output “Identified

or Unidentified”

Iris Image Ear Image



4.3 Image Preprocessing

Introduction of an image pre-processing step can radically reduce error rates. It has also been

observed that different image pre-processing techniques influence each subsequent method

differently. The results produced by some image processing techniques can be unfavorable e.g.,

blurring, smoothing, hue representations and comprehensive normalization, and others are

generally beneficial e.g., sharpen, detail, edge, enhance. Some image preprocessing techniques

may decrease error rates for some methods while increasing error rates for others.

In our proposed method, the images from the three datasets have been preprocessed by using 2-D

Gaussian low pass filter and laplacian of Gaussian filter. The results of image preprocessing are

shown in fig. 4.2 below. First the colored image is converted into a gray scale image. The gray

scale image is introduced into the Gaussian filter. This filter produces a blurred image. Then this

blurred image is passed through the laplacian of Gaussian (log) filter that will enhance the areas

of depressions and elevations. This „log‟ filter image is added to the blurred image. The resultant

image is now preprocessed and will enter to the feature extraction stage.

(a) (b) (c)

(d) (e)

Fig. 4.2: Preprocessing Results (a) Original Image, (b) Gray-scale Image, (c) Blurred Image

using Gaussian filter, (d) „Log‟ filter Image, and (e) Final Image from (c) and (d)



4.4 Feature Extraction Methods

The goal of the feature extraction method is to extract characteristic features from noisy images

that are capable of distinguishing two or more images and at the same time provide invariance

with respect to image orientation for the same image. We have used three methods for feature

extraction; DCT, LDA and PCA.

4.4.1 Implementation of DCT

DCT is incorporated into the system because of its strong energy compaction property for most

accurate feature selection. There are several other advantages offered by DCT which include:

− Provides good cooperation between computational complexity and energy packing

capability

− The energy packing property of DCT is the best among all the other unitary transforms

− Ability to pack most of the information into as fewest number of coefficients as possible

− Builds the best sub-image estimate and, thus, the smallest reconstruction errors

DCT is a data independent feature extraction method as compared to the KLT which is data

dependent. Since DCT has less computational complexity to other transforms, therefore we have

used it to extract the features of face, iris and ear images in the proposed system.

Multidimensional DCT is computed on the same lines as one-dimensional definitions: they are

simply computed along each dimension and then separable product is implemented. For

example, a two-dimensional DCT of an image or a matrix is calculated by simply finding DCT

along one dimension i.e., rows and then along the other dimension i.e., columns (or vice versa).

Mathematically, the 2-dimentional DCT is given by the formula:

∑ ∑

(

) ) [

(

) ]

4.1

∑ ∑

[

(

) ]

(

)

4.2



Eight of the total DCT coefficients for the three biometric identifiers are taken as a feature vector

representing the face image, a feature vector representing the iris image and a feature vector

representing the ear image. DCT is most efficient in terms of time complexity among the three

feature extractors being used. Its time complexity is O(n2).

4.4.2 Implementation of LDA

The key idea behind LDA is to find the subspace that is capable of distinguishing different

images by maximizing intra-class variation, while minimizing the inter-class variation. The

eigenvectors for LDA are computed by calculating the eigenvectors of Sw-1

(within class

variation) and Sb (between class variation). Here, Sb and Sw are the between-class and within-

class variation matrices defined as:

∑ ∑ ) )

4.3

∑ )

)

4.4

where ni shows the number of training samples in ith

class and mi, the mean image for ith

class.

We have employed LDA on the face, iris and ear datasets and the eigenvector of length eight is

computed. It is important to mention here that LDA feature vector takes the highest computation

time as compared to PCA and DCT methods. It is one of the drawbacks of LDA. LDA has (mnt

+ t3) time complexity and requires O(mn +mt + nt) memory, where m is the number of samples,

n is the number of features and t = min(m, n).

4.4.3 Implementation of PCA

PCA aims at finding the best set of sub-space projections in order to maximize intra-class scatter

among all the images. For this purpose, a set of Eigen-faces from the eigenvectors is computed.

The eigenvectors of the total scatter matrix St is defined as:



∑ )

)

4.5

where m represents the mean image of the sample set x. For dimensionality reduction, a subset k

(where k < m) of the eigenvectors related to first k highest Eigen-values of

St are chosen as eigenvector values.

The PCA is applied on the face, iris and ear dataset and the eigenvector of length eight is

computed. PCA uses less time as compared to LDA for computing eigenvector on the same

dataset. This is due to the fact that the time complexity of PCA, O(n3), is less than LDA.

4.5 Feature Vector Normalization

Normalization is the method of changing the location and range parameters of the score

distributions so as to transform them into a common domain. A scale parameter determines the

statistical dispersion of the probability distribution. A larger scale parameter implies a more

spread out distribution and a smaller scale parameter implies a more concentrated distribution.

For example, if two matchers had score values in the range [0, 10] and [0, 1000], the

normalization technique can be applied and the scores can be transformed into a common range

[0, 1].The location parameter determines where the origin will be located and can be either

positive or negative. The location parameter is used to shift a distribution in one direction or

another. The prediction of location and range parameters for a specific score distribution must be

efficient and robust for producing good normalization results [10]. Efficiency means the

closeness of the obtained predicted values to the optimal prediction and Robustness means the

insensitivity to outliers when the distribution is already known.

We have used z-score normalization technique. It is implemented by taking arithmetic mean and

standard deviation of the score values. It performs well when the variance and average of score

distribution are known. If this prior knowledge is not available, the mean and standard deviation

of the score values are required to be computed from given data at that time. The normalized

scores are given by

4.6



where σ represents standard deviation and µ, the arithmetic mean. For the feature vectors of our

three dataset, we have computed the average and variance first and then normalized them using

the above equation of z-score normalization.

4.6 Feature Level Fusion

The finest among the three outputs of the each identifier is selected. Now these three features are

fused together. This process is called feature level fusion. In this kind of fusion, the feature

vectors created by various biometric algorithms (DCT, PCA, and LDA) are consolidated into a

single feature vector by applying appropriate feature level fusion methods. The benefit of

feature-level fusion is finding the correlated values produced by different algorithms and thus

determining a prominent set of features that can enhance identification accuracy [18]. Ben-

Yacoub et al. [29] developed a bimodal biometric system using face and voice biometric traits.

They have explained the use of several fusion strategies, such as multilayer perceptrons, support

vector machines (SVM) and tree classifiers.

The fused feature vector can be generated by combining the best feature vectors of face, iris and

ear among the three feature extraction systems and doing feature selection on the concatenated

vector. Also, the feature sets being fused reside in commensurate vector space. Probabilistic rules

have been used which are product rule, min rule, and sum rule.

− Product rule (mul rule) performs best as described by [19] in situation when unlike

biometric traits of a user (e.g. face, ear, iris) are independent of each other. When input

feature vectors Z is entered into the matching algorithm I (i = 1, 2, 3... M)

where M shows total number of classifiers, the equation generates extracted feature

vector, xi. Wj stands for the class j (j = 1, 2, 3... m), where m shows total number of

classes. P (wj|xi) stands for the posteriori probability of the input feature vectors Z that

belongs to class wj, given the feature vector xi. Input feature vectors Z is finally allocated

to the class c that belongs to 1, 2... m.

∏ ( | )

4.7



− Min rule can produce degraded performance in the presence of noise (i.e., outliers). Then

the input feature vectors Z is allocated to class c such that

( | )

4.8

− Sum Rule produces good fusion results as it is a smoothing operation. The input feature

vectors Z is assigned to class c such that

∑ ( | )

4.9

4.7 Feature Vector Optimization using GA

Next the feature vector is now passed into the GA for optimization that repeatedly processes the

fused vector and selects most suitable features from the fused feature vector and discards the

others.

4.7.1 Genetic Algorithm

A genetic algorithm (or short GA) is a search operation employed to find fairly accurate solution

to optimization problems. They belong to a specific class of evolutionary algorithms that have

designed its techniques motivated by the concepts of evolutionary biology such as mutation,

selection, inheritance, crossover, and recombination.

Algorithm is started with some vector set of solutions which is called the initial population of the

algorithm. The set of solution is represented by chromosomes. The solutions of one population

are obtained and are utilized in formation of a new population. This is encouraged to reveal the

fact that the new population will be capable of producing better results than the old one. Those

solution sets, which are selected to generate new offspring, are selected according to their fitness

(designed by the fitness function) – New solutions (offspring) have more chances to reproduce if

they are more appropriate.

This process is repeated again and again until some stopping criteria, (for example specified

number of generations) is met.



The figure below describes in detail the various steps involve in processing genetic algorithm:

Fig. 4.3 Genetic Algorithm process flow

We have implemented genetic algorithm on the multimodal system gained after fusion of

biometric traits. The feature vector obtained by three fusion functions is then incorporated in the

genetic algorithm in order to produce more accurate scores for classification. The parameters

selected for the algorithms are described in the table 4.1.

Table 4.1: Parameters of Genetic Algorithm

4.8 Classification

using Bayesian Classifier

Parameter

Value

Population instances 100

Generations 50

Fitness Scaling Rank

Selection Roulette

Mutation Uniform

Crossover Heuristic

Start

Generation of Feature

Vector

Rank Fitness Scaling

Roulette Parent

Selection

Heuristic Crossover

Uniform Mutation

Rank fitness scaling

of children

New generation

Generations

> 50?

Optimized Feature

vector



Now the selected GA vector is introduced into the classifier that matches the GA vector to all the

vectors in the database. Therefore Bayesian classifier is incorporated which gives the output by

comparing the result from GA optimized output with the database. The system finally suggests

that the person (to be identified) is “identified” or “Unidentified”. This step completes the

identification process of our proposed system.

4.8.1 Bayesian Classifier

Bayesian classifiers are used for classification in many applications as they offer many

advantages. They are based on probability theory. They can take into account both the expert

opinion and data at the same time to construct models. They offer backward reasoning

(prediction of inputs when outputs are given), in addition to forward reasoning (prediction of

outputs when inputs are given). They also provide support when data is missing during learning

as well as classification. Bayes theorem establishes a connection between the probabilities of A

and B, denoted here by P(A) and P(B), and the conditional probabilities of A on B and B on A,

P(A|B) and P(B|A) [42]. Mathematically, the theorem is implemented by the following equation:

| ) | ) )

)

4.10

Summary

This chapter discusses our proposed method in order to solve the problems of selection of feature

extraction method and feature vector optimization, described in the last chapter. DCT has been

proposed as a feature extraction method because of its energy compaction property and reduced

computation time. The old techniques of feature extraction like PCA and LDA have also been

employed at feature extraction level in order to have a comparison between the proposed and the

existing methods.

Before the implementation of fusion methods it was necessary that the feature vectors must be

normalized. For this reason, z-score normalization is done on each of the feature vectors of face,

iris and ear trait. Z-score normalization arranges the data to a common range for fusion in next

level. The normalized feature vectors so obtained are then combined using sum, min and

product. At this level three multimodal systems have been achieved based on DCT, PCA and

http://en.wikipedia.org/wiki/Marginal_probability

http://en.wikipedia.org/wiki/Conditional_probability



LDA. Next Genetic Algorithm (GA) is employed for feature vector optimization to enhance the

efficiency of the three multimodal systems. For classification Bayesian classifier has been

implemented using supervised learning. The classifier confirms the identity of the person.

Chapter 5 Experimental Results





5.1 Introduction

Experiments have been performed to determine the accuracy of our proposed method. This

chapter discusses various scenarios for performing experiments at four different levels. At first,

the unimodal biometric traits are compared after the feature extraction to test the efficiency of

each biometric trait. Secondly, the fusion methods of „sum‟, „min‟ and „mul‟ for the three

biometric traits of face, iris and ear have been compared to get the effectiveness of the

multimodal biometric system. Next the feature extraction methods of DCT, LDA and PCA are

compared for both unimodal and multimodal systems. Also the use of genetic algorithm for

better performance of the multimodal system has been analyzed. Finally the accuracy is

computed for the three feature extraction methods with and without the implementation of

genetic algorithm.

5.2 Datasets

We have used face, iris and ear biometric traits for our multimodal biometric system. In a multi-

biometric system, it is possible sometimes that the database employed is not the true database.

True database is the one which has more than one biometric trait for each individual. Instead, the

database is a virtual database which contains records that are created by properly pairing a user

from one unimodal database (e.g., face) with a user from another database (e.g., iris). Due to the

unavailability of the biometric traits for a single person, we have implemented such a database

by pairing the three sets of face, iris and ear datasets. In this way, a triplet set of 100 is formed

from three datasets (face, iris, and ear) or we can say that a dataset of 100 individuals with face,

iris, and ear images has been used for performing our evaluation according to the proposed

methodology given in last chapter. The details of the datasets are listed as under:

5.2.1 AMI Ear database [50]

It consists of ear images collected from students, teachers and official staff at ULPGC, Spain.

The images have been taken in an indoor environment. The database was acquired from 100

different subjects, all of them in the age range of 19-65 years.



The database was acquired from 100 different subjects and 7 images per individual.

Size of the images is 492 x 702, having JPEG file format

Among seven images of each individual, there is 1 image of left ear and 6 images of right

ear with orientations: up, down, left, right and zoom

5.2.2 Faces94 Dataset [49]

A sequence of 20 images is taken when a subject is asked to speak while sitting at a fixed

distance from the camera. The speech introduces variation in facial expressions. Database

description is as under:

Total number of subjects is 153; female (20), male (113), male staff(20)

Image resolution is 180 × 200 pixels with JPEG format

The detail of variation in individual‟s images is given below:

The background is plain green

Very minor variations seen in the attributes like head in turn, tilt and slant position

lighting variation is not introduced while taking the images

Considerable facial expression changes as speech has been introduced

5.2.3 MMU2 Iris Database [48]

MMU2 iris database has a total of 995 iris images. These images were taken using Panasonic

BM-ET100US camera with an operating range of 47-53 cm. These iris images were taken from

volunteers with different nationality and age. They are natives of Europe, Asia, Africa and

Middle East.

The database was made from 100 different subjects and 10 images per individual

5 images each of left and right eye have been included

Size of the images is 320 × 238, having Bitmap file format



5.3 Performance Measures

The following performance measures have been used to estimate the effectiveness of our

proposed method [51]:

False Accept Rate (FAR): It is a measure of the likelihood that the biometric system

wrongly accepts the input user to a non-matching user template in the enrollment

database. It measures the percentage of impostor users which are wrongly accepted.

F R

6.1

False reject rate (FRR): It is a measure of the likelihood that the biometric system is

unable to find a match between the input user and a matching user template in the

enrollment database. It measures the percentage of genuine users which are wrongly

rejected.

6.2

Receiver operating characteristic (ROC): The ROC plot is a visual display of the graph

between the FAR and the FRR. A threshold determines proximity of the input to a

template in order to consider the input as a match. If the value of the threshold is

decreased, the FAR increases and FRR reduces. Similarly, a higher threshold will

decrease the FAR and at the same time increases the FRR. The closer the ROC curves to

the origin, the better the system.

Fig. 5.1: ROC Curve

http://en.wikipedia.org/wiki/Receiver_operating_characteristic



Equal error rate (EER): It is the value obtained from the ROC curve where the values of

FAR and FRR are equal. The EER attempts to exploit the accuracy of different biometric

systems using the ROC curves. In general, the lower the EER, the more accurate the

results.

{

) )

) ) ) )

) )

6.3

Where | ) ) and | ) ) .

Accuracy (ACC): the accuracy is the proportion of true results i.e., both the number of

genuine accepted and number of imposters rejected in the population. It is a parameter of

the test.

6.4

5.4 Experimental Results

5.4.1 Results of Unimodal Biometrics

The three set of images (face, iris and ear) are introduced to each of the three feature extraction

models (DCT, LDA and PCA). In this way, three feature scores are obtained for each of the three

models e.g. face score, iris score, and ear score for DCT model and similarly for the other two

models. At this level, the scores are compared to find out three things: Which biometric trait is

more efficient in identification of a person? Which feature extraction model outperforms the

other models?

The EER of each feature extraction method of each database is presented in Table 5.1. We see

that the proposed method of DCT produces results that are better than the existing methods in

case of iris and ear biometric traits. However the best result is shown by LDA using face



biometric trait with the least error. At this level, we are only comparing unimodal biometric

system and the existing methods do perform better for unimodal systems but in case of

multimodal system (that is demonstrated next) these existing methods cannot perform well. In

order to reduce spoof attacks, implementation of multimodal system is preferred.

Table 5.1: Performance (EER) of three feature extraction methods for face iris and ear biometric

traits

Feature

Extraction

Method

Error Equal Rate

(EER)

Face Iris Ear

PCA 0.5542 0.5083 0.4333

LDA 0.3458 0.5750 0.5083

DCT 0.5333 0.4667 0.4000

Figure 5.2 shows the ROC curves of feature extraction methods implemented on the face, iris

and ear datasets separately as individual biometric systems. Here the ear features have the best

EER value in case DCT because the closer the graph to the origin the better the performance of

the biometric trait. LDA is giving better performance in case of face features and PCA curve is

showing overlap of all the traits, ear being the better.

It has been seen that each biometric trait has its advantages and limitations, and a single trait can

never meet all the requirements effectively such as efficiency, practicality and expenses at the

same time. Therefore, we can say that there is no universally best accepted biometric trait; search

for best biometric trait is still going on. Also the selection of particular biometric trait depends on

the nature and requirements of the particular application for which the trait is to be employed.



(a) (b)

(c)

Fig. 5.2: ROC performance curve of face, iris and ear biometric traits using (a) DCT, (b) LDA,

and (c) PCA

5.4.2 Results of Multimodal Biometrics

Three fusion techniques („sum‟, „min‟ and „mul‟) are used for combining the feature scores of

the biometric traits. This is done for finding the efficiency of our proposed multimodal biometric

system. As we have used logistic regression method, so weights to the three biometric traits have

been assigned; 0.3, 0.4 and 0.3 are assigned to face, iris and ear respectively. A lower value of

weight is assigned to the more accurate biometric trait (the face and ear matcher being more



accurate here). Therefore, the results here can be expected to be more influenced by the ranks

assigned by the face and ear scores which is the case.

Table 5.2 gives a comparison of the performance measure (EER) of these fusion techniques for

the feature extraction methods with and without using genetic algorithm. We see that the genetic

algorithm has reduced the EER value to a much greater extent. However, the proposed DCT

method is giving accurate results in both cases (with and without GA). The best results with

lowest EER are produced by DCT with „mul‟ fusion technique.

Table 5.2: Performance (EER) of three fusion techniques for PCA, LDA and DCT

Feature Extraction

Method

Error Equal Rate

(EER)

Sum Min Mul

Without GA

PCA 0.5083 0.5083 0.4875

LDA 0.5083 0.4625 0.4625

DCT 0.3792 0.3667 0.4125

With GA

PCA 0.0643 0.0952 0.2929

LDA 0.2310 0.7381 0.5000

DCT 0.0524 0.0214 0.0143

Figure 5.3 shows the performance of two multimodal biometric systems: one without genetic

algorithm and the other with genetic algorithm. The fusion techniques (sum, min and mul) have

been compared for each feature extraction method (DCT, LDA, and PCA). Figure 5.3 (b, e)

gives the comparison between two LDA systems one without genetic algorithm and the other

with genetic algorithm, respectively. The straight lines in (b) show that the imaginary values

accounted by LDA have not been properly calculated. This problem has been overcome by still



having low EER values. This defect has greatly been reduced by the implementation of genetic

algorithm on this method as shown by (e).

(a) (b)

(c) (d)

(e) (f)



Fig. 5.3: ROC performance curves of „sum‟, „min‟ and „mul‟ fusion techniques using (a) DCT,

(b) LDA, (c) PCA, (d) DCT with GA, (e) LDA with GA, and (f) PCA with GA

The figure 5.4 shows the comparison between best fusion techniques in all feature extraction

methods. We can see that genetic algorithm reduces the EER value in case of each method.

Using DCT the „mul‟ fusion technique performs the best when genetic algorithm is employed as

shown in (a). Similarly using LD and PC , „min‟ and „mul‟ fusion techniques are producing

good results respectively. The straight line in (b) showing LDA curve becomes straight as the

curve proceeds downwards. This is because here the complex values are not accurately evaluated

and effected the results badly. However, the error rate is still very low so this defect can be

neglected.

(a) (b)



(c)

Fig. 5.4: ROC Performance curves of Best Fusion technique in terms of EER with and without

the use of Genetic Algorithm using (a) DCT, (b) LDA, and (c) PCA

5.4.3 Results of Feature Extraction Methods

Figure 5.5 shows a comparison between the best feature extraction methods on the basis of

„sum‟, „min‟ and „mul‟ fusion techniques. The EER as well as the ROC show better performance

of DCT over the other two methods among all the fusion techniques.

(a) (b)

(c)

Fig. 5.5: ROC Performance curves of feature extraction methods using (a) „sum‟, (b) „min‟ and (c) „mul‟



We then selected the best fusion methods in each of the feature extraction methods with the

lowest EER values; The DCT and LD with „mul‟, and the PC with „min‟. They are compared

by the ROC curves as shown in the figure 5.6 below. DCT is producing better results here as

well. LDA and PCA are overlapping but still PCA produced better results here than LDA.

Fig. 5.6: Comparison of the Best fusion techniques of the three feature extraction methods

To our knowledge, we can say now that the DCT with genetic algorithm has produced the better

results as individual biometric and as multimodal biometric systems. However multimodal

biometric systems always perform better than individual biometric systems.

The feature extraction methods are also compared in terms of time complexity. For our database

with 100 images each of face, iris and ear, making a total of 300 images, the DCT is using less

time as compared to the other existing methods of LDA and PCA as shown in figure 5.7. This is

due to the fact that the DCT is far less computationally complex than PCA and LDA. The



complexity of 2D-DCT is O(n2) whereas for PCA and LDA it is O(n

3) and O(mnt+t

3)

respectively where m is the number of samples, n is the number of features and t = min(m, n).

Table 5.3: Time Consumption by feature Extraction methods

Feature Extraction

Method

Time

Complexity

Time (seconds) used

by 300 images

PCA O(n3) 16.793765

LDA O(mnt+t3) 27.393744

DCT O(n2) 8.794472

5.4.4 Classification Accuracy Measurement

The multimodal systems have been classified by using Naive Bayesian Classification Model.

The accuracy has been computed by the confusion matrix produced in each case of fusion

technique before and after implementing genetic algorithm as shown in figure 5.8. In each case

the accuracy of the multimodal system has been greatly enhanced by the use of genetic

algorithm. The classification accuracy values of simple multimodal system and GA multimodal

system are listed in table 5.3.

In case of „sum‟ and „mul‟, the DCT and LD perform well as compared to PC whereas in

case of „min‟, PC outperforms LD . So with different fusion methods, PC and LD are

performing differently. However, DCT performs better in all cases irrespective of the fusion

technique being implemented as shown by the bar graphs in figure 5.8.



Table 5.4: Accuracy of the simple multimodal systems and the GA multimodal systems

Feature Extraction

Method

Accuracy

(%)

Sum Min Mul

Without GA

PCA 90.00 89.89 85.00

LDA 92.94 87.50 89.74

DCT 93.33 90.00 94.11

With GA

PCA 93.75 91.91 87.00

LDA 94.11 88.75 92.30

DCT 95.55 92.85 96.47

(a) (b)

87

88

89

90

91

92

93

94

95

96

DCT LDA PCA

Acc

ura

cy (

%)

Without GA

With GA

84

85

86

87

88

89

90

91

92

93

DCT LDA PCA

Acc

ura

cy (

%)

Without GA

With GA



(c)

Fig. 5.7: Bar graph showing accuracy of fusion techniques with and without the use of Genetic Algorithm

for (a) sum, (b) min, and (c) mul

Summary

An illustration of experiments being performed to calculate the efficiency and accuracy of the

proposed method is described. The chapter discusses different scenarios for performing

experiments at four different levels. At first, the unimodal biometric traits are compared after the

feature extraction to test the efficiency of each biometric trait. Here it has been seen that ear and

face biometric traits are performing well as compared to iris. For this reason, when these

biometric traits are fused for making a multimodal system the weight of iris is set higher than

face and ear trait. Higher weight is set for those traits that produce less accurate results. In this

way, contribution of iris trait towards multibiometric results is reduced.

Secondly, the fusion methods of „sum‟, „min‟ and „mul‟ for the three biometric traits of face, iris

and ear have been compared to get the effectiveness of the multimodal biometric system. Next

the feature extraction methods of DCT, LDA and PCA are compared for both unimodal and

78

80

82

84

86

88

90

92

94

96

98

DCT LDA PCA

Acc

ura

cy (

%)

Without GA

With GA



multimodal systems. Also the use of genetic algorithm for better performance of the multimodal

system has been analyzed. Finally the accuracy is computed for the three feature extraction

methods with and without the implementation of genetic algorithm. The final classification

results have shown that the accuracy of LDA and PCA depends on the fusion method being used,

as in case of „min‟, PC has shown improved accuracy than LD . In sum and product fusion

methods, LDA shows enhanced performance than PCA. However, the accuracy of the proposed

DCT method has outperformed LDA and PCA in all case of fusion methods.

Chapter 6 Conclusion and Future Work





6.1 Introduction

Efficient identity management system has become very important in this highly interconnected

world with increased concerns of identity fraud and national security. Biometric systems provide

a greater degree of security and user convenience than the traditional authentication methods.

Moreover, biometric systems also provide negative recognition and non-repudiation that

traditional systems don‟t. Multibiometric systems, if properly designed, are able to increase the

matching accuracy of a recognition system as they, consolidate the evidences from different

biometrics, increase population coverage and prevent spoofing attacks.

6.1 Conclusion

In this thesis, we have developed a new approach to enhance the recognition accuracy of

proposed multibiometric system. The system uses Discrete Cosine Transform at feature

extraction level due to its high energy compaction property and reduced computation time. The

existing methods of Principal Component Analysis and Linear Discriminant Analysis are also

used in parallel in order to compare the three feature extraction methods. Three biometric traits

are used; face iris and ear. It has been seen that each biometric trait has its own advantages and

limitations, and a single trait can never meet all the necessary requirements effectively such as

efficiency, practicality and expenses at the same time. Therefore, we can say that there is no

universally best accepted biometric trait; search for best biometric trait is still going on. Also the

selection of particular biometric trait depends on the nature of the particular application for

which the trait is to be employed.

The results prove the increased performance of DCT over the other state of the art methods.

Three biometric traits i.e. face, iris and ear have been used which at fused at feature level to

make the system a multimodal biometric system. ROC curves in terms of FAR and FRR has

shown the better performance of our proposed DCT method for feature extraction.

It has been seen that after normalization of the feature vector, the different fusion techniques

implemented on the three systems (DCT, LDA and PCA) perform differently. So we cannot

summarize that which fusion method performs well for all systems. Genetic algorithm

incorporated for feature vector optimization has produced greater accuracy as compared to the



simple (without GA) methods; the highest classification accuracy being achieved is 96.47% by

the DCT feature extraction method using „mul‟ fusion technique. The use of Discrete Cosine

Transform with genetic algorithm has significantly improved the performance of our

multibiometric system.

6.2 Future Work

The accuracy of biometric technology depends on the accuracy and number of records within the

databases. Therefore if the data capturing module is not efficient, which is most of the case, then

required accuracy of identification cannot be met. So a lot of work is still required in the area of

multibiometrics to reduce the effects of noisy data and to increase template security.

The work presented in this thesis can be extended:

− By implementing different variation of PCA and LDA like Multilinear PCA, kernel PCA,

Independent Discriminant Analysis, etc.

− By using the energy compaction property of DCT on a feature extraction method of

LDA, and PCA. In this way the feature vector of LDA and PCA systems would be

optimized by DCT.

− By employing various fusion techniques in parallel and concatenating the resultant

vectors.

− By implementing fusion at various levels described in chapter 2 for making a comparison

on the multibiometric system. For example, fusion at the rank level can be implemented

and the methods of logistic regression, borda count and highest rank can be compared.

References


References

[1] L. Hong and A.K. Jain, " Integrating Faces and Fingerprints For Personal

Identification", IEEE Transactions Pattern Analysis and Machine Intelligence , Vol.20,

No.12, pp 1295-1307, 1998

[2] R. Frischholz, U. Dieckmann,“BiolD: multimodal biometric identification system”,

Computer, Vol. 33,. No. 2, pp 64-68, 2000

[3] M. Martinez, . C. Kak, “PC versus LD ”. IEEE Transactions on Pattern nalysis and

Machine Intelligence , Vol.23, No.2, pp 228-233, 2001

[4] J. Fierrez-Aguilar, J. Ortega-Garcia, D. Garcia-Romero, J. Gonzalez Rodriguez, “

comparative evaluation of fusion strategies for multimodal biometric verification”, 4th

International Conference Audio-video-based Biometric Person Authentication, pp 830-837,

2003

[5] A. Kumar, D. C. M. Wong, H. C.Shen, . K. Jain, “Personal verification using palm print

and hand geometry biometric”, 4th International Conference udio-Video-Based Biometric

Person Authentication, pp 668-678, 2003

[6] T. Heseltine, N. Pears, J. ustin, Z. Chen, “Face recognition: comparison based on

appearance-based approaches”, 7th Digital Image Computing: Techniques and applications,

2003

[7] K. . Toh, X. D. Jiang, W. Y. Yau, “Exploiting global and local decisions for multi-modal

biometrics verification”, IEEE Trans. Signal Process., pp 3059–3072, 2004

[8] R. Snelick, U. Uludag, A. Mink, M. Indovina, and A. Jain, " Large Scale Evaluation of

Multimodal Biometric Authentication Using State-of-the-Art Systems", IEEE Transactions

on Pattern Analysis and Machine Intelligence , Vol. 27, No. 3, pp. 450-455, March 2005

[9] . Ross, “ n introduction to multimodal biometrics”, 15th European Signal Processing

Conference (EUSIPCO), 2007

[10] . Kumar, Y. Zhou, “Human Identification Using KnuckleCodes,” IEEE International

Conference on Biometrics: Theory; Applications and Systems, Vol.45, No. 20, pp.1-7,

September 2009

References


[11] S. Hocquet, J.Y. Ramel, H. Cardot, “User classification for keystroke dynamics

authentication”, The Sixth International Conference on Biometrics (ICB2007), 2007

[12] P. Teh, . Teoh, T. Ong, H. Neo, “Statistical Fusion pproach on Keystroke Dynamics,”

Third International IEEE Conference on Signal Image Technologies and Internet-Based

System; IEEE Computer Society, 2007

[13] T.C. Sabareeswari, S.L. Stuwart, “Identification of a Person using Multimodal Biometric

System” International Journal of Computer pplications, 2010

[14] Giot, M. El- bed, C. Rosenberger “Fast Learning For Multibiometrics Systems Using

Genetic lgorithms” Workshop on Security and High Performance Computing Systems

(SHPCS); The IEEE International Conference on High Performance Computing &

Simulation (HPCS), 2010

[15] A. Patra, Development of Efficient Methods for Face Recognition and Multimodal

Biometry” MS Thesis submitted to Indian Institute of Technology Madras, 2006

[16] J. D. Woodward, C. Horn, J. Gatune, and A. Thomas. Biometrics: A look at Facial

Recognition. RAND Documented Briefing, Tech. Rep., 2003.

[17] Ross, ., Nandakumar, K. and Jain, .K., “Handbook of Multibiometrics”, Springer-

Science + Business Media, LLC, 2006.

[18] “Biometrics”, available at http://en.wikipedia.org/wiki/Biometrics

[19] P. Dhamla, “Multibiometric Systems”, MS Thesis submitted to Norwegian University of

Science and Technology, 2012

[20] Z. M. Hafed, M. D. Levine, “Face Recognition using Discrete Cosine Transform”,

International Journal of Computer Vision, 2001

[21] K. Nandakumar, “Multibiometric Systems: Fusion Strategies and Template Security, PhD

Thesis submitted to Michigan State University, 2008

[22] “Feature Extraction”, available at http://en.wikipedia.org/wiki/Feature_extraction

[23] S. . Khayyam, “The Discrete Cosine Transform: Theory and pplication”, Seminar 1 –

Information Theory and Coding, Michigan State University, 2003

http://en.wikipedia.org/wiki/Biometrics

http://en.wikipedia.org/wiki/Feature_extraction

References


[24] R. S. Choras, “Image Feature Extraction Techniques and Their Applications for CBIR and

Biometrics Systems”, International Journal of Biology and Biomedical Engineering”, 2007

[25] M. Pechenizkiy, S. Puuronen, Tsymbal, “The impact of sample reduction on PC based

feature extraction for supervised learning”, Proceedings of the 2006 CM symposium on

Applied computing, 2006

[26] S. S. Iyengar, L. Prasad, and H. Min. Advances in Distributed Sensor Technology Prentice

Hall, 1995.

[27] Ross, ., K. Nandakumar, .K. Jain, “Handbook of Multibiometrics”, Springer-Science +

Business Media LLC, 2006

[28] J. Soh, F. Deravi , . Triglia, “Multibiometrics and data fusion standardization”, in

Encyclopedia of Biometrics Springer, 2009

[29] S. BenYacoub, Y. bdeljaoued, E. Mayoraz, “Fusion of Face and Speech Data for Person

Identity Verification”, IEEE Transactions on Neural Networks, 1999

[30] K. . Toh, W. Xiong, W. Y. Yau, X. Jiang, “Combining Fingerprint and Hand-Geometry

Verification Decisions”, Fourth International Conference on Audio and Video based

Biometric Person Authentication (AVBPA), 2003

[31] E. S. Bigun, J. Bigun, B. Duc, S. Fischer, “Expert Conciliation for Multimodal Person

uthentication Systems using Bayesian Statistics”, First International Conference on Audio-

and Video-Based Biometric Person Authentication (AVBPA), 1997

[32] R. Snelick, U. Uludag, . Mink, M. Indovina, . K. Jain, “Large Scale Evaluation of

Multimodal Biometric Authentication Using State-of-the-Art Systems”, IEEE Transactions

on Pattern Analysis and Machine Intelligence, 2005

[33] R. Brunelli, D. Falavigna, “Person Identification Using Multiple Cues”, IEEE Transactions

on Pattern Analysis and Machine Intelligence, 1995

[34] . K. Jain, L. Hong, Y. Kulkarni, “ Multimodal Biometric System using Fingerprint, Face

and Speech, “Second International Conference on udio and Video based Biometric Person

Authentication (AVBPA), 1999.

References


[35] J. Kittler, M. Hatef, R. P. Duin, J. G. Matas, “On Combining Classifiers”, IEEE

Transactions on Pattern Analysis and Machine Intelligence, 1998

[36] S. Krawczyk, . K. Jain, “Securing Electronic Medical Records using Biometric

uthentication”, Fifth International Conference on udio and Video based Biometric Person

Authentication (AVBPA), 2005

[37] Y. Wang, T. Tan, A. K. Jain, “Combining Face and Iris Biometrics for Identity

Verification”, Fourth International Conference on Audio and Video based Biometric Person

Authentication (AVBPA), 2003

[38] L. Hong, . K. Jain, “Integrating Faces and Fingerprints for Personal Identification”, IEEE

Transactions on Pattern Analysis and Machine Intelligence, 1998

[39] K. Chang, K. W. Bowyer, S. Sarkar, B. Victor, “Comparison and Combination of Ear and

Face Images in Appearance-based Biometrics”, IEEE Transactions on Pattern Analysis and

Machine Intelligence, 2003

[40] “Bayes' theorem”, available at http://en.wikipedia.org/wiki/Bayes%27_theorem

[41] G. Feng, K. Dong, D. Hu, D. Zhang, “When Faces are Combined with Palm prints: A Novel

Biometric Fusion Strategy”, First International Conference on Biometric Authentication

(ICBA), 2004

[42] K. . Toh, X. Jiang, W. Y. Yau, “Exploiting Global and Local Decisions for Multimodal

Biometrics Verification”, IEEE Transactions on Signal Processing, 2004

[43] K. . Toh, W. Y. Yau, “Fingerprint and Speaker Verification Decisions Fusion Using a

Functional Link Network”, IEEE Transactions on Systems: Man and Cybernetics, 2005

[44] S. K. Sahoo, S. R. M. Prasanna, S. R. T. Choubisa, “Multimodal Person uthentication:

Review”, IETE Technical Review, 2012

[45] J. Fierrez-Aguilar, J. Ortega-Garcia, J. Gonzalez-Rodriguez, J. Bigun, “Discriminative

Multimodal Biometric uthentication based on Quality Measures”, Pattern Recognition,

2005

References


[46] G. Shakhnarovich, L. Lee, T.J. Darrell, “Integrated Face and Gait Recognition from

Multiple Views”, IEEE Conference on Computer Vision and Pattern Recognition (CVPR),

2001

[47] . Ross, . K. Jain, “Information Fusion in Biometrics” Pattern Recognition Letters, 2003

[48] “MMU2 Iris Database”, available at http://pesona.mmu.edu.my/~ccteo/

[49] “Faces94 Face Database by Libor Spacek”, available at

http://cswww.essex.ac.uk/mv/allfaces/faces94.html

[50] “ MI Ear Database by Esther Gonzalez, Luis Alvarez and Luis Mazorra”, available at

http://www.ctim.es/research_works/ami_ear_database/

[51] A. Kale, A. K. Roy Chowdhury, R. Chellappa, “Fusion of Gait and Face for Human

Identification”, IEEE International Conference on coustics, Speech, and Signal Processing

(ICASSP), 2004

[52] J. Jacob, N. T. Bhuvan, S. M. Thampi, “India Feature Level Fusion using Multiple

Fingerprints”, IJC Special Issue on Computational Science, New Dimensions &

Perspectives (NCCSE), 2011

http://pesona.mmu.edu.my/~ccteo/

http://cswww.essex.ac.uk/mv/allfaces/faces94.html



Appendix

Multimodal Biometric Security using Evolutionary Computation i

Glossary

This section lists acronyms that frequently appear in the thesis.

DCT Discrete Cosine Transform

LDA Linear Discriminant Analysis

PCA Principal Component Analysis

GA Genetic Algorithm

ROC Receiver Operating Characteristic

EER Error Equal Rate

PIN Personal Identity Number

ROI Region of Interest

DNA De-oxyribo Nucleic Acid

FAR False Accept Rate

FRR False Reject Rate

IR Infra Red

SVM Support Vector Machines

HMM Hidden Markov Model

ACC Accuracy

KLT Karhunen Loeve Transform

JPEG Joint Photographic Experts Group (image file format)

PNG Portable Network Graphics (image file format)

RGB Red Green Blue

COTS Counter-of-the-Shelf

FTE Fail To Enroll

FMR False Match Rate

https://en.wikipedia.org/wiki/Karhunen%E2%80%93Lo%C3%A8ve_theorem

https://en.wikipedia.org/wiki/Joint_Photographic_Experts_Group

Appendix

Multimodal Biometric Security using Evolutionary Computation ii

FNMRFalse Non-Match Rate

DWT Discrete Wavelet Transform

PDA Personal Digital Assistant

MCC Multiple Classifier Combination

FTER Failure to Enroll Rate

FTCR Failure to Capture Rate

JVT Joint Video Team

Documents

Multimodal Biometric Security using Evolutionary Computation · 2016-09-09 · Multimodal Biometric Security using Evolutionary Computation 2 1.1 Introduction to Biometric Systems